Finding defects it was never shown
Three surfaces that vary from part to part, five kinds of small, subtle defects. Each detector learns from 120 good parts per surface only and is graded on 140 test parts with pixel-exact masks.
0.974image AUROC, mean over surfaces (pixel statistics 0.847, PCA 0.764)
77%of defects caught with 3.9% of good parts flagged
0.950pixel AUROC: the map points at the defect, not just the part
43 msper image on one CPU core, no GPU and no pretrained network
Image AUROC by surface
| method | weave | tiles | wood | mean |
|---|---|---|---|---|
| pixel statistics | 0.914 | 0.850 | 0.778 | 0.847 |
| PCA reconstruction | 0.778 | 0.792 | 0.722 | 0.764 |
| patch NN, random bank | 0.957 | 0.925 | 0.910 | 0.931 |
| patch nearest neighbour | 0.990 | 0.981 | 0.950 | 0.974 |
Image AUROC by defect
| method | scratch | spot | hole | crack | stain |
|---|---|---|---|---|---|
| pixel statistics | 0.797 | 0.955 | 1.000 | 0.668 | 0.817 |
| PCA reconstruction | 0.697 | 0.885 | 0.996 | 0.593 | 0.649 |
| patch NN, random bank | 0.974 | 0.957 | 1.000 | 0.964 | 0.759 |
| patch nearest neighbour | 0.998 | 0.990 | 1.000 | 0.997 | 0.883 |
At the operating threshold (set so 2% of separate good parts are flagged)
| method | defects caught | good parts flagged | ms per image |
|---|---|---|---|
| pixel statistics | 44.2% | 0.0% | 0.5 |
| PCA reconstruction | 35.0% | 2.8% | 1.1 |
| patch NN, random bank | 60.0% | 3.3% | 35.9 |
| patch nearest neighbour | 77.1% | 3.9% | 42.6 |
Examples (top: the part with the defect outlined; bottom: the anomaly map)
weave, scratch
tiles, crack
wood, spot
weave, stain
tiles, hole